HarnessSentry – least-privilege auditor for AI coding harnesses

Help developers see and shrink what Claude Code and Codex agents can touch on their machine

Problem

Harnesses like Claude Code and Codex have broad access to the local machine and its stored secrets, and the risk compounds once MCPs and scheduled API tasks are involved. Most developers have no clear picture of which files, env vars, and credentials an agent can actually reach.

Solution

A local-first permission auditor that scans harness configs, MCP manifests, and scheduled jobs, then lists every file, shell capability, env var, and secret the agent can reach. It suggests fixes and generates a hardened, sandboxed launch config.

Users

Individual developers and small teams running AI coding agents heavily, especially those juggling multiple agents with MCPs and API credentials

MVP

Scan common Claude Code / Codex config paths; flag high-risk secrets and writable directories; generate a least-privilege launch script or container config; produce a shareable risk report

Revenue

Open-source core plus paid team tier with policy templates, CI integration, and scheduled scan alerts, priced per seat

Why now

Agents are moving from chat windows to operating the machine directly, and permission boundaries just became a visible pain; enterprise security tools ignore this personal-dev niche